Structural finding
Loops, exact cycles, retry storms, tool churn, stagnation, and structural health tiers from T1-T5. T4 remains review-only; T5 is critical structural evidence.
CAUM Agent Waste Audit
Find where agents burn compute without clean structural progress.
Send CAUM your agent traces. We return loops, retry storms, stagnation, token and cost exposure, and a remediation receipt that engineering teams can replay without exposing private content.
Zero-semantic
Observe-only
No truth scoring
No blocking
No guaranteed savings claim
What the audit returns
Detection is only useful when it becomes an engineering decision.
CAUM turns traces into a review package: what happened structurally, where cost was exposed, what should be changed, and what evidence supports the recommendation.
Cost exposure
Token and cost counters tied to the suspicious segment. CAUM reports reviewable exposure, not guaranteed savings and not a public waste percentage.
Fix receipt
Recommended interventions such as retry ceilings, checkpoints, summarize-and-reset, tool budgets, exit conditions, or human review gates.
Sample Fix Receipts
Two real validation receipts: one comparison run, one controlled loop.
These are May 1, 2026 CAUM Live production-validation receipts. They show the product shape we sell: structural finding, observed counters, recommended intervention, and claim boundaries.
Receipt A
No hard alert
Comparison run: production website smoke.
A non-control browser-agent workflow produced review-only structural pressure, but no exact cycle, no hard alert, and no autonomous control recommendation.
T3
Review-only tier
0.0
Exact-cycle coverage
$0.002
Observed cost
- Finding
- Structural pressure observation. This is a watch signal, not a confirmed loop and not a public waste rate.
- Evidence
- 5 core steps, 0 cycle intervals, 0.0 exact-cycle coverage, 170 observed tokens, allowed_to_block=false.
- Action
- Continue monitoring. No loop intervention recommended from this receipt alone.
Receipt B
T5 hard alert
Controlled retry loop: missing-route retry control.
A controlled non-destructive validation run entered a repeated two-state browser cycle and produced critical structural evidence.
1.0
Exact-cycle coverage
1,356
Observed tokens
$0.0155
Reviewable exposure
- Finding
- Critical structural loop: 12 core steps, 1 exact cycle interval, 6 repeated cycles, period 2, 2 unique states.
- Evidence
- T5 public signal, hard_alert=true, strong_exact_cycle=true, critical_t5_evidence=true, observed cost $0.018.
- Action
- Add a retry ceiling, route-existence precheck, and checkpoint reset after repeated two-state browser cycles. Rerun the workflow and compare cycle coverage.
Claim lock: these receipts show structural evidence and reviewable cost exposure only. They do not score answer truth, predict future outcomes, block agents, require private content, or publish a universal waste rate.
Before / After Validation
CAUM becomes more valuable when detection turns into repair.
On May 2, 2026, the same missing-route control was rerun with a retry guard pattern. CAUM Live observed the before/after structure without reading private content or controlling the agent.
Before
Repeated retry loop.
Public signal
T5 hard alert
Exact-cycle coverage
1.0
Observed tokens
1,356
Reviewable exposure
$0.015459
After
Retry guard applied.
Public signal
T2 observation
Exact-cycle coverage
0.0
Observed tokens
281
Reviewable exposure
$0.000393
The intervention CAUM recommends.
This is a remediation-pattern validation, not a universal savings guarantee. It shows how a structural finding can become an engineering change and a replayable receipt.
- Add a route-existence precheck before repeating the same browser action.
- Stop after the first missing-route result instead of re-entering the same two-state cycle.
- Fallback to a known route and write a checkpoint that records the retry guard decision.
- Rerun representative workflows and compare hard-alert status, exact-cycle coverage, tokens, and reviewable exposure.
Boundary: this validates a remediation pattern on a controlled non-destructive workflow. It does not claim CAUM guarantees savings, predicts success, judges answer truth, or controls the agent.
When the receipt repeats, move to CAUM Live.
The PDF receipt is the proof artifact. CAUM Live is the operating layer that watches recurring agent workflows and keeps structural exposure visible while work is happening.
- Use one PDF receipt to prove the loop, retry, or stagnation pattern.
- Connect CAUM Live to the recurring workflow once the pattern appears more than once.
- Track hard alerts, exact-cycle coverage, tokens, and reviewable exposure across active runs.
- Use fix receipts to tune retry ceilings, checkpoints, tool budgets, and review gates.
Why this is the commercial entry point.
Small teams need a cheap receipt. Larger teams need a defensible audit. CAUM can serve both without pretending every customer has enterprise-scale spend.
1,350public traces replayed
54,100structural events observed
5.3%hard-alert traces
0production replay errors
1
Send traces without secretsUpload JSON, JSONL, or stream operational events. Prompts, files, customer messages, and private outputs can stay out.
2
CAUM produces structural evidenceHealth tier, exact-cycle coverage, hard-alert status, token exposure, cost exposure, and claim-audited boundaries.
3
Engineering gets a remediation receiptConcrete changes to retry policy, checkpointing, tool budgets, or review gates that can be tested against the same workflow.
Start at the right size
One ladder, not one price for everyone.
CAUM should be cheap for builders, paid for teams, and serious only when the agent spend or operational risk is serious.
Builder
Trace Receipt
$99 / report target
- After-the-run trace audit
- Loops, retries, stagnation, cost exposure
- Best for small API spend or debugging
Best first sale
Team Waste Audit
$2,500 / initial audit target
- 7-14 days of agent traces
- Structural waste findings and fix receipts
- Conservative executive summary
Live
Enterprise Pilot
$10k+ / scoped pilot target
- Live structural monitoring
- Multiple agent workflows
- Governance-safe, observe-only integration
Public claims stay conservative: CAUM reports structural evidence and reviewable cost exposure. It does not score answer truth, guarantee savings, or control agent behavior.
The next concrete move
Start with a receipt. Keep value with Live.
The fastest path to revenue is one trace receipt, then one team audit, then CAUM Live on the workflows where loops and retry exposure repeat.